Source: 
SRResNetVgg5,4: 
(Ground Truth: 
)
PyTorch version of the paper: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (currently it does not implement the GAN, but the srresnet + vgg19-5,4 loss)
you can train a net from scratch:
(optionally start training with just the pixel-wise loss on the resnet part:
python srresnet.py --image-dir traindir --cuda --pretraining --images 16384 --batchSize 16)
(use --pretrained modelfile.pth to continue from a pretraining or previous run for example)
python srresnet.py --image-dir traindir --cuda --images 16384 --batchSize 16
and then inference with the arguments:
--pretrained model/model_epoch_80.pth --testing --test-image BSDS300/images/train/100075.jpg